position_stack() should stack +ive & -ve values separately

[hadley]

library(ggplot2)
library(tibble)

df <- frame_data(
  ~x, ~g, ~y,
  1,  1,  1,
  1,  2,  -1,
  1,  3,  1,
  2,  1,  2,
  2,  2,  -3
)

ggplot(df, aes(x, y, fill= factor(g))) + 
  geom_bar(stat = "identity")

[thomasp85]

There are a lot of case where, even if supported, negative values doesn't make sense when stacking. As I see it it only makes sense for stacked bar charts - can you come with other examples...

I'm asking because I need to know which types of checks to make and warnings to throw, and also where it should be solved. It could be handled directly by pos_stack() but this function is also called by pos_fill() where it doesn't make sense to have negative values. Alternatively (and easier) the data could be split into two calls to collide in PositionStack$compute_panel() which would make the functionality specific to PositionStack (though it would still be a mess if coupled with geom_area

[hadley]

Can you come with an example where stacking makes sense apart from bars/areas? I think that's the only case we need to worry about, and it's reasonable for positive values to stack upwards from the y-axis, and negative values to stack downwards.

Your fix seems reasonable to me - I don't think there's any way to have an elegant solution without completely rewriting the stacking (which we don't want to do)

[thomasp85]

I cannot - which was why I asked whether you knew of any other use cases. I'll move forward with #1693

[smouksassi]

I think we can also stack dots which can be useful in some cases. http://www.lispworks.com/success-stories/inspiredata.html

http://docs.ggplot2.org/current/geom_dotplot.html

[thomasp85]

@smouksassi true. Dot plots should mainly be used for count data though, which makes them unsuitable to show negative values.

[smouksassi]

I was thinking about a kind of back to back dotplot like this: https://nzmaths.co.nz/resource/fully-stretched:

[hadley]

Dot plots have to use a completely different algorithm so aren't relevant to this discussion.

[wmurphyrd]

Here's an example where the current behavior of position_stack for a mixture of positive and negative values produced the desired result: the total extent of the bar aligns with the sum of the values (please ignore the politics, just examples that I happened to have handy).
. However, I did zoom-hide the negative side because it's not clear what's going on. (From https://github.com/wmurphyrd/BernieTax )

In another example, I wanted to show what was happening on the negative side with a dodge:
. Although, in the end I used geom_tile with position_dodge after calculating the offsets manually

[thomasp85]

I do think that these examples, while potentially useful to you, violates the general expectations of stacking, namely that stacks are build from the ground and up (or down), that is, they extend from zero

[hadley]

Cool plots, and it is a nice use, but I think it also violates the principle that you expect position_stack() to avoid overlaps (i.e. even in this case you want to dodge as well to show the direction of each change).